Skip to content

Cache-Aside — Middle

At middle level, focus on this question:

How does the TTL you choose trade off staleness against database load?

Prerequisite: junior.md.

TTL controls the staleness/load trade-off

flowchart LR Short["Short TTL (10s)"] --> S1["Fresh data,\nmore cache misses,\nmore DB load"] Long["Long TTL (1 hour)"] --> L1["Fewer DB queries,\nbut data can be up to\n1 hour stale"]

Every cached entry expires after its TTL, forcing the next request for that key to miss and re-fetch from the database — this is the only mechanism that corrects staleness in plain cache-aside (no invalidation), so the TTL is a direct, tunable bound on "how wrong can this cached value be, at worst."

TTL choice Good for
Seconds Rapidly-changing data (stock prices, live counters) where staleness is expensive
Minutes to hours Data that changes occasionally (user profile, product catalog)
Very long / no expiry Reference data that almost never changes (country codes, static config) — pair with explicit invalidation (07-cache-invalidation/) instead of relying on TTL alone

Miss vs. stale: two different costs

A cache miss costs one database round trip — bounded, predictable. Serving stale data (a value that's technically still cached but no longer matches the database) costs something much harder to bound: a wrong answer served to a user or downstream system, for up to the TTL's full duration.

def get_user(user_id):
    cached = cache.get(f"user:{user_id}")
    if cached is not None:
        return cached   # could be up to TTL seconds stale - is that OK here?
    ...

🎓 Takeaway: choosing a TTL is choosing how much staleness your specific use case can tolerate, in exchange for how much database load you're willing to accept. There's no universal "correct" TTL — it's a per-key-type business decision.

Test yourself

  1. Why is a 1-hour TTL potentially dangerous for a "current inventory count" cache key, but perfectly fine for a "country list" cache key?
  2. If database load spikes right after a deploy, and you suspect it's from cache misses, what would you check first?
  3. Design two different TTLs for two different fields on the same user object — which field gets the shorter TTL, and why?

Continue to senior.md.